The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes external data in the form of JSON tour scenarios and test result artifacts, which could contain instructions intended to influence the agent's behavior.
Ingestion points: External JSON files located in dogfood/scenarios/ and media/log artifacts fetched from the pr-media branch.
Boundary markers: The instructions do not specify any delimiters or safety warnings for the agent when processing these external files.
Capability inventory: The skill has extensive system capabilities, including running local Python/Shell scripts (run-e2e.sh, verify-local.py), performing ssh connections, and executing build commands (xcodebuild, swiftc).
Sanitization: There is no mention of sanitization or validation of the instructions contained within the JSON tours before execution.
[DYNAMIC_EXECUTION]: The ui-lab.py utility compiles Swift source code at runtime to generate UI previews.
Evidence: scripts/ui-lab/ui-lab.py is described as compiling a harness plus app source files using swiftc to render PNGs.
[COMMAND_EXECUTION]: The skill relies on executing various repository-local scripts and system binaries to perform verification tasks.
Evidence: Commands include python3 scripts/verify-local.py, scripts/ui-test, scripts/run-e2e.sh, and xcodebuild test.

cmux-testing